The weekend left one big governance story and a stack of agent news that is easy to miss. Nvidia’s board opened an inquiry into its own chief executive, Microsoft showed what attackers are doing with agents, and Amazon and CrowdStrike both shipped tools that put agents to work or stop them cold.
1. Nvidia’s Board Opens an Independent Probe Into Jensen Huang Reports
What happened. On Nov 8, 2026, Nvidia’s board said it is launching an independent investigation after reports that founder and CEO Jensen Huang was the target of a foreign intelligence effort. A committee of independent directors will run it with outside counsel, and the board said Huang supports the review and keeps its confidence.
The hype check. The statement is short on specifics: it does not name the directors on the committee, the law firm, or a timeline, and it does not describe what the reports allege. Its line that the board has no reason to believe national security or company secrets were compromised is a judgment made before the investigation has done its work, so treat it as a starting position rather than a finding.
Verdict: News. The investigation is confirmed by the board itself, and the outcome is the part worth waiting for.
2. Microsoft Says Attackers Use Agents to Turn Leaked Keys Into Breaches
What happened. Microsoft Threat Intelligence reported that an agentic attack framework called ARTEX has been searching public GitHub repositories for exposed secrets since August 2026, including API keys for major AI providers. In one case Microsoft watched an agent go from a GitHub search to pulling data out of an Azure Key Vault in about five minutes. A follow-up post on Nov 6 says a Cambodia-based criminal group Microsoft tracks as Storm-3358 has used the same framework against financial firms since October.
The hype check. Microsoft does not say who builds ARTEX, does not attribute the GitHub campaign to a named actor, and gives no full victim count. The advice is old advice made urgent: a key that was ever public should be treated as stolen and rotated, even if the commit was deleted. It echoes CrowdStrike’s earlier report on hackers running DeepSeek agents against hundreds of targets.
Verdict: Supported. Microsoft published detailed observations and indicators from its own telemetry, though the people behind the tool are still unnamed.
3. Nvidia Courts Korea With Startup Stakes and Model Serving Recipes
What happened. Nvidia used a forum with Korea’s science ministry to announce an investment in interconnect maker Point2 and plans to invest in Motif Technologies, RLWRLD and Upstage. It also released Dynamo deployment recipes for four Korean open models, including LG’s K-EXAONE 2.0 and SK Telecom’s A.X K2, and published a 7 million record synthetic Korean persona dataset on Hugging Face.
The hype check. No investment amounts were disclosed, and three of the four stakes are described as planned. The recipes promise up to twice the performance out of the box on B200 GPUs, but Nvidia names no baseline or benchmark. The wider sovereign AI collaboration only lays groundwork for Korean chip makers to explore NVLink Fusion. For builders, the open dataset is the most concrete piece, much like Nvidia’s recent free Nemotron speech model.
Verdict: Promising but unverified. The speed claim lacks a stated baseline and the deal sizes are not public.
4. Amazon’s Shopping Agent Now Takes Orders by Text Message
What happened. Amazon says customers can now text Alexa for Shopping to find products, compare options, build a cart and buy, including recurring grocery orders and searches from a photo. Conversations carry over to the Amazon app, the website and Alexa devices.
The hype check. For now it is invitation only, for a small group of U.S. customers, with wider access promised in the coming weeks and no date. Amazon’s page says reorders happen with your approval but does not spell out a confirmation step for every purchase. Agent shopping has been slow to catch on elsewhere, as DoorDash’s count of only several hundred agent orders showed.
Verdict: News. The feature is real but tiny so far, and how many people use it is still unknown.
5. CrowdStrike’s Agent Kill Switch Works at the Gateway First
What happened. CrowdStrike introduced an AI Kill Switch in Falcon Guardian that lets an authorized admin stop a running agent, block it from relaunching on chosen hosts, and cut its network access, with each step logged. The company says the feature is available now.
The hype check. Today the stop is enforced at CrowdStrike’s gateway by cutting the agent off from the models it calls, while enforcement on the endpoint itself is listed as coming soon. The release gives no pricing, no list of supported agent types, and no outside test of how fast or complete the shutdown is. It fits the identity and control questions in NIST’s design tenets for agent identities.
Verdict: Needs follow-up. Half of the promised control is still to ship, and nobody outside the company has tested it.
Also on our radar
- A Tencent Hunyuan and Zhejiang University team built SplatAnything, a 3D scene reconstruction method that scales scenes by the spacing between points and holds up better when scene size or camera angle changes.
- The TailSteer paper on multi-turn image editing adds a short extra training stage that cuts the visual drift editors pile up over 15 to 20 rounds of changes.
- Researchers proposed EvoMemState, a memory system for personal agents that tracks whether each stored fact is still valid and beat the best baselines on its benchmark by 6 to 12 points.
- A position paper on cognitive symbionts argues that AI progress should be judged by how much it strengthens a person over a lifetime, not by how much it does alone.
What to watch
Watch for Nvidia to name the law firm and the directors on its committee, and for any sign the inquiry affects its leadership or government relationships. On the security side, the real test is whether agent attacks like ARTEX push companies to rotate keys and adopt controls such as kill switches faster than attackers can automate. Amazon’s promised expansion of text shopping to more customers should show whether people trust an agent with their cart.
Sources
- NVIDIA Newsroom: Statement from NVIDIA Board of Directors
- Microsoft Security Blog: Threat actors use agentic AI tooling to find and abuse exposed secrets
- Microsoft Security Blog: ARTEX abuse expands as cybercriminals target financial sector
- NVIDIA Blog: NVIDIA Announces New Planned Investments in Korean Startups at the MSIT-NVIDIA AI Startup Partnership Forum
- NVIDIA Blog: NVIDIA Dynamo Recipes Bring Korea’s Leading Open Models to Production
- NVIDIA Blog: Korea’s AI Ecosystem Advances and Secures Agentic AI With NVIDIA
- NVIDIA Blog: From Chips to Skills, NVIDIA and Korea Expand Sovereign AI Collaboration
- Amazon: Text, shop, done: Alexa for Shopping is now just a text away
- CrowdStrike: CrowdStrike Introduces AI Kill Switch for AI Safety
- Hugging Face Papers: SplatAnything: Gaussian Spacing Normalization for Feed-forward 3DGS across Arbitrary Scenes
- Hugging Face Papers: TailSteer: Trajectory-Tail Steering for Drift-Resistant Multi-turn Image Editing
- Hugging Face Papers: EvoMemState: Evolving Memory State for Personal Agents
- Hugging Face Papers: From Autonomous Agents to Cognitive Symbionts: Toward AI-Augmented Human Intelligence
By Willa Ravenwire · The AI Bandwagon
